2024 workshop program
PulsarNet - Accelerating the Discovery of Binary Pulsars through Attention-Based Neural Networks
The work
Abstract
In the realm of radio astronomy, the discovery and analysis of binary pulsars in compact orbits offer a unique opportunity to test theories of gravity, particularly General Relativity, in the strong-field limit. In this talk, I will introduce PulsarNet, a machine learning (ML) pipeline designed for the detection and parameter estimation of binary pulsars in time-series observations. PulsarNet is the first fully ML-based search pipeline of its kind, marking a significant departure from conventional matched filtering techniques. It achieves comparable sensitivity and offers substantially faster processing speeds than traditional searches. In the second half, I will discuss the European Research Council (ERC)-funded project “COMPACT,” which I am currently involved with. COMPACT will collect 1.6 PB of raw baseband data every two months from the MeerKAT telescope, targeting Globular clusters. I will provide an overview of the state-of-the-art algorithms used for binary pulsar searching within this project and the unique challenges encountered in processing PB-scale data. Additionally, I will share insights on how machine learning has enabled us to discover over 78 pulsars in the ongoing quasi-real-time pulsar searches in the MPIfR MeerKAT Galactic Plane Survey.